Wouldn't attention networks use less computation for the same number of weights? Feed forward networks have higher connectivity no? If a network is the same si

Wouldn't attention networks use less computation for the same number of weights? Feed forward networks have higher connectivity no? If a network is the same size, less computationally demanding, and gives you 3% improvement, it seems extremely worthwhile, especially given the (mostly) diminishing returns of just adding more weights/layers.